base.py 32.8 KB
Newer Older
Leo Gao's avatar
Leo Gao committed
1
import abc
2
from typing import Iterable
thefazzer's avatar
thefazzer committed
3
import numpy as np
4
import random
Leo Gao's avatar
Leo Gao committed
5
import re
6
7
8
import os
import json
import hashlib
Jonathan Tow's avatar
Jonathan Tow committed
9
import datasets
10
from sqlitedict import SqliteDict
11
from tqdm import tqdm
12
import torch
Leo Gao's avatar
Leo Gao committed
13
import torch.nn.functional as F
14
from accelerate import find_executable_batch_size
&'s avatar
& committed
15

16
from lm_eval.metrics import mean, weighted_perplexity, weighted_mean, bits_per_byte
17
from lm_eval import utils
18
from abc import abstractmethod
Jason Phang's avatar
gpt3  
Jason Phang committed
19

Jason Phang's avatar
Jason Phang committed
20

Leo Gao's avatar
Leo Gao committed
21
class LM(abc.ABC):
Leo Gao's avatar
Leo Gao committed
22
23
24
    def __init__(self):
        self.cache_hook = CacheHook(None)

25
    @abstractmethod
Leo Gao's avatar
Leo Gao committed
26
    def loglikelihood(self, requests):
Leo Gao's avatar
Leo Gao committed
27
        """Compute log-likelihood of generating a continuation from a context.
Fabrizio Milo's avatar
Fabrizio Milo committed
28
        Downstream tasks should attempt to use loglikelihood instead of other
Leo Gao's avatar
Leo Gao committed
29
        LM calls whenever possible.
Jason Phang's avatar
gpt3  
Jason Phang committed
30

Leo Gao's avatar
Leo Gao committed
31
32
33
        :param requests: list
            A list of pairs (context, continuation)
            context: str
Fabrizio Milo's avatar
Fabrizio Milo committed
34
                Context string. Implementations of LM must be able to handle an
Leo Gao's avatar
Leo Gao committed
35
                empty context string.
Leo Gao's avatar
Leo Gao committed
36
            continuation: str
Fabrizio Milo's avatar
Fabrizio Milo committed
37
38
                The continuation over which log likelihood will be calculated. If
                there is a word boundary, the space should be in the continuation.
Leo Gao's avatar
Leo Gao committed
39
40
41
42
                For example, context="hello" continuation=" world" is correct.
        :return: list
            A list of pairs (logprob, isgreedy)
            logprob: float
Jason Phang's avatar
Jason Phang committed
43
                The log probability of `continuation`
Leo Gao's avatar
Leo Gao committed
44
            isgreedy:
Jason Phang's avatar
Jason Phang committed
45
46
47
48
                Whether `continuation` would be generated by greedy sampling from `context`
        """
        pass

49
    @abstractmethod
Leo Gao's avatar
Leo Gao committed
50
    def loglikelihood_rolling(self, requests):
Jason Phang's avatar
Jason Phang committed
51
52
53
54
        """Compute full log-likelihood of a string, with no truncation, for perplexity computation
        - We will use the full max context length of the model.
        - For inputs that exceed the max context length, we divide the tokenized string into chunks of up to
        the max context length.
Fabrizio Milo's avatar
Fabrizio Milo committed
55
        - IMPORTANT: Each document's loglikelihood/perplexity is computed *separately*, unlike other implementations
Jason Phang's avatar
Jason Phang committed
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
          which may simply concatenate multiple documents together.
        - IMPORTANT: We maximize the amount of context for each prediction. Specifically, for inputs that we break into
          multiple chunks, the last input will still a full-sized context.
          Example:
            Input tokens: [ 0 1 2 3 4 5 6 7 8 9 ]
            Prefix: EOT
            Max context length: 4
            Resulting input/prediction pairs:

                INPUT:  EOT   0   1   2
                PRED:     0   1   2   3

                INPUT:    3   4   5   6
                PRED:     4   5   6   7

                INPUT:    5   6   7   8
                PRED:             8   9

          Observe that:
            1. Each token is predicted exactly once
            2. For the last pair, we provide the full context, but only score the last two tokens

        :param requests: list
            A list of strings
            string: str
                String for which we are computing per-toke  loglikelihood
        :return: list
            A list of pairs (logprob, isgreedy)
            logprob: float
                The log probability of `continuation`
            isgreedy:
                Whether `continuation` would be generated by greedy sampling from `context`
Leo Gao's avatar
Leo Gao committed
88
89
90
        """
        pass

&'s avatar
& committed
91
    # TODO: Add an optional max length
92
    @abstractmethod
Leo Gao's avatar
Update  
Leo Gao committed
93
    def greedy_until(self, requests):
Leo Gao's avatar
Leo Gao committed
94
95
96
97
98
99
        """Generate greedily until a stopping sequence

        :param requests: list
            A list of pairs (context, until)
            context: str
                Context string
Leo Gao's avatar
Leo Gao committed
100
            until: [str]
Fabrizio Milo's avatar
Fabrizio Milo committed
101
                The string sequences to generate until. These string sequences
Leo Gao's avatar
Leo Gao committed
102
                may each span across multiple tokens, or may be part of one token.
Leo Gao's avatar
Leo Gao committed
103
104
105
106
        :return: list
            A list of strings continuation
            continuation: str
                The generated continuation.
Jason Phang's avatar
gpt3  
Jason Phang committed
107
        """
Leo Gao's avatar
Leo Gao committed
108
109
        pass

Jason Phang's avatar
gpt3  
Jason Phang committed
110
    @classmethod
111
112
    def create_from_arg_string(cls, arg_string, additional_config=None):
        additional_config = {} if additional_config is None else additional_config
113
114
115
        args = utils.simple_parse_args_string(arg_string)
        args2 = {k: v for k, v in additional_config.items() if v is not None}
        return cls(**args, **args2)
Jason Phang's avatar
gpt3  
Jason Phang committed
116

Leo Gao's avatar
Leo Gao committed
117
118
119
    def set_cache_hook(self, cache_hook):
        self.cache_hook = cache_hook

Leo Gao's avatar
Leo Gao committed
120

121
class BaseLM(LM):
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
    @property
    @abstractmethod
    def eot_token_id(self):
        pass

    @property
    @abstractmethod
    def max_length(self):
        pass

    @property
    @abstractmethod
    def max_gen_toks(self):
        pass

    @property
    @abstractmethod
    def batch_size(self):
        pass

    @property
    @abstractmethod
    def device(self):
        pass

147
    @abstractmethod
Fabrizio Milo's avatar
Fabrizio Milo committed
148
149
150
    def tok_encode(self, string: str):
        pass

151
    @abstractmethod
Fabrizio Milo's avatar
Fabrizio Milo committed
152
153
    def tok_decode(self, tokens: Iterable[int]):
        pass
Jason Phang's avatar
gpt3  
Jason Phang committed
154

155
    @abstractmethod
Fabrizio Milo's avatar
Fabrizio Milo committed
156
157
    def _model_generate(self, context, max_length, eos_token_id):
        pass
Jason Phang's avatar
gpt3  
Jason Phang committed
158

159
160
    @abstractmethod
    def _model_call(self, inps):
Jason Phang's avatar
gpt3  
Jason Phang committed
161
        """
162
163
        inps: a torch tensor of shape [batch, sequence]
        the size of sequence may vary from call to call
Jason Phang's avatar
gpt3  
Jason Phang committed
164

165
        returns: a torch tensor of shape [batch, sequence, vocab] with the
166
        logits returned from the model
167
168
        """
        pass
169

Leo Gao's avatar
Leo Gao committed
170
    # subclass must implement properties vocab_size, eot_token_id, max_gen_toks, batch_size, device, max_length.
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
    # TODO: enforce this somehow

    def loglikelihood(self, requests):
        new_reqs = []
        for context, continuation in requests:
            if context == "":
                # end of text as context
                context_enc = [self.eot_token_id]
            else:
                context_enc = self.tok_encode(context)

            continuation_enc = self.tok_encode(continuation)

            new_reqs.append(((context, continuation), context_enc, continuation_enc))

        return self._loglikelihood_tokens(new_reqs)

    def loglikelihood_rolling(self, requests):
        # TODO: Implement caching once we've confirmed the perplexity implementation
190

191
192
        # automatic batch size detection for vectorization
        adaptive_batch_size = None
193
        if self.batch_size == "auto":
194
            # using rolling window with maximum context
195
196
197
198
199
            print("Passed argument batch_size = auto. Detecting largest batch size")

            @find_executable_batch_size(
                starting_batch_size=512
            )  # if OOM, then halves batch_size and tries again
200
            def forward_batch(batch_size):
201
202
203
204
205
                test_batch = torch.ones(
                    (batch_size, self.max_length), device=self.device
                ).long()
                for _ in range(5):
                    _ = F.log_softmax(self._model_call(test_batch), dim=-1).cpu()
206
                return batch_size
207
208

            batch_size = forward_batch()
209
210
            print(f"Determined Largest batch size: {batch_size}")
            adaptive_batch_size = batch_size
211
212

        loglikelihoods = []
Fabrizio Milo's avatar
Fabrizio Milo committed
213
214
215
216
217
218
219
220
221
222
223
224
        for (string,) in tqdm(requests):
            rolling_token_windows = list(
                map(
                    utils.make_disjoint_window,
                    utils.get_rolling_token_windows(
                        token_list=self.tok_encode(string),
                        prefix_token=self.eot_token_id,
                        max_seq_len=self.max_length,
                        context_len=1,
                    ),
                )
            )
225
226
227

            rolling_token_windows = [(None,) + x for x in rolling_token_windows]

228
229
            # TODO: extract out this call so it only gets called once and also somehow figure out partial caching for
            # that
Fabrizio Milo's avatar
Fabrizio Milo committed
230
            string_nll = self._loglikelihood_tokens(
231
232
233
                rolling_token_windows,
                disable_tqdm=True,
                override_bs=adaptive_batch_size,
Fabrizio Milo's avatar
Fabrizio Milo committed
234
235
            )

236
237
            # discard is_greedy
            string_nll = [x[0] for x in string_nll]
Fabrizio Milo's avatar
Fabrizio Milo committed
238

239
240
241
242
243
            string_nll = sum(string_nll)
            loglikelihoods.append(string_nll)

        return loglikelihoods

244
    def _loglikelihood_tokens(self, requests, disable_tqdm=False, override_bs=None):
245
246
247
248
249
250
        # TODO: implement some kind of efficient-request-middleware that lumps together requests with the same context
        res = []

        def _collate(x):
            # the negative sign on len(toks) sorts descending - this has a few advantages:
            # - time estimates will always be over not underestimates, which is more useful for planning
251
252
253
            # - to know the size of a batch when going through the list, you know the first one is always the batch
            #   padded context length. this is useful to simplify the batching logic and more importantly to make
            #   automatic adaptive batches much much easier to implement
254
255
256
            # - any OOMs will happen right away rather than near the end

            toks = x[1] + x[2]
257
            return -len(toks), tuple(toks)
Fabrizio Milo's avatar
Fabrizio Milo committed
258

Fabrizio Milo's avatar
Fabrizio Milo committed
259
        re_ord = utils.Reorderer(requests, _collate)
260
261
262

        # automatic (variable) batch size detection for vectorization
        # pull longest context sample from request
263
264
265
266
        if len(re_ord.get_reordered()) > 0:
            _, context_enc, continuation_enc = re_ord.get_reordered()[0]
            max_context = len((context_enc + continuation_enc)[-(self.max_length + 1) :][:-1])
            if (self.batch_size == 'auto'):
267
                
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
                if override_bs is None:
                    print('Passed argument batch_size = auto. Detecting largest batch size')
                    @find_executable_batch_size(starting_batch_size=512) # if OOM, then halves batch_size and tries again
                    def forward_batch(batch_size):
                        test_batch = torch.ones((batch_size, max_context), device=self.device).long()
                        for _ in range(5):
                            out = F.log_softmax(self._model_call(test_batch), dim = -1).cpu()
                        return batch_size

                    batch_size = forward_batch()
                    print(f"Determined largest batch size: {batch_size}")
                    adaptive_batch_size = batch_size

                else:
                    adaptive_batch_size = override_bs
        else:
            adaptive_batch_size = 0 if override_bs is None else override_bs
285

Fabrizio Milo's avatar
Fabrizio Milo committed
286
        for chunk in utils.chunks(
287
288
            tqdm(re_ord.get_reordered(), disable=disable_tqdm),
            self.batch_size if self.batch_size != "auto" else adaptive_batch_size,
Fabrizio Milo's avatar
Fabrizio Milo committed
289
        ):
290
            inps = []
291
            cont_toks_list = []
292
            inplens = []
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307

            padding_length = None

            # because vectorizing is annoying, we first convert each (context, continuation) pair to padded
            # tensors, then we pack them together into a batch, call the model, and then pick it all apart
            # again because vectorizing is annoying

            for _, context_enc, continuation_enc in chunk:
                # sanity check
                assert len(context_enc) > 0
                assert len(continuation_enc) > 0
                assert len(continuation_enc) <= self.max_length

                # how this all works:
                #          CTX      CONT
308
                # inp    0 1 2 3|4 5 6 7 8 9   <- last token is deleted by inp[:, :-1]
309
                # gpt2    \               \
310
311
                # logits   1 2 3|4 5 6 7 8 9   <- the ctx half gets tossed out by the
                # cont_toks      4 5 6 7 8 9      [:, -len(continuation_enc):, :self.vocab_size] slice
312
313
314

                # when too long to fit in context, truncate from the left
                inp = torch.tensor(
Fabrizio Milo's avatar
Fabrizio Milo committed
315
316
                    (context_enc + continuation_enc)[-(self.max_length + 1) :][:-1],
                    dtype=torch.long,
317
                ).to(self.device)
Fabrizio Milo's avatar
Fabrizio Milo committed
318
                (inplen,) = inp.shape
319
320
321
322

                cont = continuation_enc

                # since in _collate we make sure length is descending, the longest is always the first one.
Fabrizio Milo's avatar
Fabrizio Milo committed
323
324
325
                padding_length = (
                    padding_length if padding_length is not None else inplen
                )
326

327
                # pad length from seq to padding_length
Fabrizio Milo's avatar
Fabrizio Milo committed
328
329
330
331
332
333
334
335
336
                inp = torch.cat(
                    [
                        inp,  # [seq]
                        torch.zeros(padding_length - inplen, dtype=torch.long).to(
                            inp.device
                        ),  # [padding_length - seq]
                    ],
                    dim=0,
                )
337

338
339
                inps.append(inp.unsqueeze(0))  # [1, padding_length]
                cont_toks_list.append(cont)
340
                inplens.append(inplen)
341

342
            batched_inps = torch.cat(inps, dim=0)  # [batch, padding_length]
Fabrizio Milo's avatar
Fabrizio Milo committed
343
344
345
            multi_logits = F.log_softmax(
                self._model_call(batched_inps), dim=-1
            ).cpu()  # [batch, padding_length, vocab]
346

347
348
            for (cache_key, _, _), logits, inp, inplen, cont_toks in zip(
                chunk, multi_logits, inps, inplens, cont_toks_list
Fabrizio Milo's avatar
Fabrizio Milo committed
349
            ):
350

351
352
                # Slice to original seq length
                contlen = len(cont_toks)
353
                inplen = inplen + (logits.shape[0] - padding_length)
Fabrizio Milo's avatar
Fabrizio Milo committed
354
355
356
                logits = logits[inplen - contlen : inplen].unsqueeze(
                    0
                )  # [1, seq, vocab]
357

358
                # Check if per-token argmax is exactly equal to continuation
359
                greedy_tokens = logits.argmax(dim=-1)
Fabrizio Milo's avatar
Fabrizio Milo committed
360
361
362
                cont_toks = torch.tensor(cont_toks, dtype=torch.long).unsqueeze(
                    0
                )  # [1, seq]
363
364
                max_equal = (greedy_tokens == cont_toks).all()

365
366
                # Obtain log-probs at the corresponding continuation token indices
                # last_token_slice = logits[:, -1, :].squeeze(0).tolist()
Fabrizio Milo's avatar
Fabrizio Milo committed
367
368
369
                logits = torch.gather(logits, 2, cont_toks.unsqueeze(-1)).squeeze(
                    -1
                )  # [1, seq]
370

371
                # Answer: (log prob, is-exact-match)
372
373
374
375
376
377
378
379
                answer = (float(logits.sum()), bool(max_equal))

                # partial caching
                if cache_key is not None:
                    self.cache_hook.add_partial("loglikelihood", cache_key, answer)

                res.append(answer)

Fabrizio Milo's avatar
Fabrizio Milo committed
380
        return re_ord.get_original(res)
Fabrizio Milo's avatar
Fabrizio Milo committed
381

382
    def greedy_until(self, requests):
Fabrizio Milo's avatar
Fabrizio Milo committed
383
        # TODO: implement fully general `until` that handles until that are
384
        #       multiple tokens or that span multiple tokens correctly
385
386
387
388
389
390

        # TODO: extract to TokenizedLM?
        res = []

        def _collate(x):
            toks = self.tok_encode(x[0])
391
            return len(toks), x[0]
Fabrizio Milo's avatar
Fabrizio Milo committed
392

Fabrizio Milo's avatar
Fabrizio Milo committed
393
        re_ord = utils.Reorderer(requests, _collate)
394

395
        for context, request_args in tqdm(re_ord.get_reordered()):
396
            until = request_args["until"]
397
398
            if isinstance(until, str):
                until = [until]
399

400
401
402
403
            if until:
                (primary_until,) = self.tok_encode(until[0])
            else:
                primary_until = None
404

Fabrizio Milo's avatar
Fabrizio Milo committed
405
406
407
            context_enc = torch.tensor(
                [self.tok_encode(context)[self.max_gen_toks - self.max_length :]]
            ).to(self.device)
408

409
            max_gen_tokens = min(
410
                self.max_gen_toks, request_args.get("max_length", self.max_gen_toks)
411
            )
Fabrizio Milo's avatar
Fabrizio Milo committed
412
            cont = self._model_generate(
413
                context_enc, context_enc.shape[1] + max_gen_tokens, primary_until
Fabrizio Milo's avatar
Fabrizio Milo committed
414
415
416
            )

            s = self.tok_decode(cont[0].tolist()[context_enc.shape[1] :])
417
418
419

            for term in until:
                s = s.split(term)[0]
Fabrizio Milo's avatar
Fabrizio Milo committed
420

421
422
            # partial caching
            self.cache_hook.add_partial("greedy_until", (context, until), s)
Fabrizio Milo's avatar
Fabrizio Milo committed
423

424
            res.append(s)
Fabrizio Milo's avatar
Fabrizio Milo committed
425

Fabrizio Milo's avatar
Fabrizio Milo committed
426
        return re_ord.get_original(res)
Leo Gao's avatar
Leo Gao committed
427

Leo Gao's avatar
Leo Gao committed
428

429
class Task(abc.ABC):
&'s avatar
&amp; committed
430
431
432
433
434
435
436
437
    """A task represents an entire benchmark including its dataset, problems,
    answers, and evaluation methods. See BoolQ for a simple example implementation

    A `doc` can be any python object which represents one instance of evaluation.
    This is usually a dictionary e.g.
        {"question": ..., "answer": ...} or
        {"question": ..., question, answer)
    """
Jonathan Tow's avatar
Jonathan Tow committed
438

Jon Tow's avatar
Jon Tow committed
439
440
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
Jonathan Tow's avatar
Jonathan Tow committed
441
442
443
444
445
    DATASET_PATH: str = None

    # The name of a subset within `DATASET_PATH`.
    DATASET_NAME: str = None

Jon Tow's avatar
Jon Tow committed
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
    def __init__(self, data_dir=None, cache_dir=None, download_mode=None):
        """
        :param data_dir: str
            Stores the path to a local folder containing the `Task`'s data files.
            Use this to specify the path to manually downloaded data (usually when
            the dataset is not publicly accessible).
        :param cache_dir: str
            The directory to read/write the `Task` dataset. This follows the
            HuggingFace `datasets` API with the default cache directory located at:
                `~/.cache/huggingface/datasets`
            NOTE: You can change the cache location globally for a given process
            by setting the shell environment variable, `HF_DATASETS_CACHE`,
            to another directory:
                `export HF_DATASETS_CACHE="/path/to/another/directory"`
        :param download_mode: datasets.DownloadMode
            How to treat pre-existing `Task` downloads and data.
            - `datasets.DownloadMode.REUSE_DATASET_IF_EXISTS`
                Reuse download and reuse dataset.
            - `datasets.DownloadMode.REUSE_CACHE_IF_EXISTS`
                Reuse download with fresh dataset.
            - `datasets.DownloadMode.FORCE_REDOWNLOAD`
                Fresh download and fresh dataset.
        """
        self.download(data_dir, cache_dir, download_mode)
470
        self._training_docs = None
471
        self._fewshot_docs = None
sdtblck's avatar
sdtblck committed
472

Jon Tow's avatar
Jon Tow committed
473
    def download(self, data_dir=None, cache_dir=None, download_mode=None):
Fabrizio Milo's avatar
Fabrizio Milo committed
474
        """Downloads and returns the task dataset.
Jonathan Tow's avatar
Jonathan Tow committed
475
476
        Override this method to download the dataset from a custom API.

Jon Tow's avatar
Jon Tow committed
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
        :param data_dir: str
            Stores the path to a local folder containing the `Task`'s data files.
            Use this to specify the path to manually downloaded data (usually when
            the dataset is not publicly accessible).
        :param cache_dir: str
            The directory to read/write the `Task` dataset. This follows the
            HuggingFace `datasets` API with the default cache directory located at:
                `~/.cache/huggingface/datasets`
            NOTE: You can change the cache location globally for a given process
            by setting the shell environment variable, `HF_DATASETS_CACHE`,
            to another directory:
                `export HF_DATASETS_CACHE="/path/to/another/directory"`
        :param download_mode: datasets.DownloadMode
            How to treat pre-existing `Task` downloads and data.
            - `datasets.DownloadMode.REUSE_DATASET_IF_EXISTS`
                Reuse download and reuse dataset.
            - `datasets.DownloadMode.REUSE_CACHE_IF_EXISTS`
                Reuse download with fresh dataset.
            - `datasets.DownloadMode.FORCE_REDOWNLOAD`
                Fresh download and fresh dataset.
Jonathan Tow's avatar
Jonathan Tow committed
497
498
499
500
        """
        self.dataset = datasets.load_dataset(
            path=self.DATASET_PATH,
            name=self.DATASET_NAME,
Jon Tow's avatar
Jon Tow committed
501
502
            data_dir=data_dir,
            cache_dir=cache_dir,
Fabrizio Milo's avatar
Fabrizio Milo committed
503
            download_mode=download_mode,
Jonathan Tow's avatar
Jonathan Tow committed
504
        )
sdtblck's avatar
sdtblck committed
505

506
507
508
509
    def should_decontaminate(self):
        """Whether this task supports decontamination against model training set."""
        return False

510
    @abstractmethod
511
    def has_training_docs(self):
Jason Phang's avatar
checkin  
Jason Phang committed
512
        """Whether the task has a training set"""
513
        pass
514

515
    @abstractmethod
516
    def has_validation_docs(self):
Jason Phang's avatar
checkin  
Jason Phang committed
517
518
519
        """Whether the task has a validation set"""
        pass

520
    @abstractmethod
Jason Phang's avatar
checkin  
Jason Phang committed
521
522
    def has_test_docs(self):
        """Whether the task has a test set"""
523
524
        pass

Leo Gao's avatar
Leo Gao committed
525
    def training_docs(self):
Jason Phang's avatar
checkin  
Jason Phang committed
526
527
528
529
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
Leo Gao's avatar
Leo Gao committed
530
        return []
531

Leo Gao's avatar
Leo Gao committed
532
    def validation_docs(self):
533
534
535
536
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
Leo Gao's avatar
Leo Gao committed
537
        return []
538

Leo Gao's avatar
Leo Gao committed
539
    def test_docs(self):
540
541
542
543
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
Leo Gao's avatar
Leo Gao committed
544
        return []
Leo Gao's avatar
Leo Gao committed
545

Jon Tow's avatar
Jon Tow committed
546
547
548
549
    def _process_doc(self, doc):
        """
        Override this to process (detokenize, strip, replace, etc.) individual
        documents. This can be used in a map over documents of a data split.
Jon Tow's avatar
Jon Tow committed
550
        E.g. `map(self._process_doc, self.dataset["validation"])`
Jon Tow's avatar
Jon Tow committed
551
552
553
554
555
556

        :return: dict
            The processed version of the specified `doc`.
        """
        return doc

557
    def fewshot_examples(self, k, rnd):
558
559
        if self._training_docs is None:
            self._training_docs = list(self.training_docs())
560

Leo Gao's avatar
Leo Gao committed
561
        return rnd.sample(self._training_docs, k)
Leo Gao's avatar
Leo Gao committed
562

563
    def doc_to_decontamination_query(self, doc):
Fabrizio Milo's avatar
Fabrizio Milo committed
564
565
566
567
        print(
            "Override doc_to_decontamination_query with document specific decontamination query."
        )
        assert False
568

569
    @abstractmethod
Leo Gao's avatar
Update  
Leo Gao committed
570
571
572
    def doc_to_text(self, doc):
        pass

573
    @abstractmethod
Leo Gao's avatar
Update  
Leo Gao committed
574
    def doc_to_target(self, doc):
Leo Gao's avatar
Leo Gao committed
575
        pass
Leo Gao's avatar
Leo Gao committed
576

577
    @abstractmethod
578
    def construct_requests(self, doc, ctx):
Fabrizio Milo's avatar
Fabrizio Milo committed
579
        """Uses RequestFactory to construct Requests and returns an iterable of
Leo Gao's avatar
Leo Gao committed
580
581
        Requests which will be sent to the LM.

582
583
        :param doc:
            The document as returned from training_docs, validation_docs, or test_docs.
Leo Gao's avatar
Leo Gao committed
584
        :param ctx: str
Fabrizio Milo's avatar
Fabrizio Milo committed
585
            The context string, generated by fewshot_context. This includes the natural
586
            language description, as well as the few shot examples, and the question
Fabrizio Milo's avatar
Fabrizio Milo committed
587
            part of the document for `doc`.
Leo Gao's avatar
Leo Gao committed
588
        """
Leo Gao's avatar
Leo Gao committed
589
        pass
590

591
    @abstractmethod
Leo Gao's avatar
Leo Gao committed
592
    def process_results(self, doc, results):
Fabrizio Milo's avatar
Fabrizio Milo committed
593
594
        """Take a single document and the LM results and evaluates, returning a
        dict where keys are the names of submetrics and values are the values of
595
        the metric for that one document
Leo Gao's avatar
Leo Gao committed
596
597
598
599
600

        :param doc:
            The document as returned from training_docs, validation_docs, or test_docs.
        :param results:
            The results of the requests created in construct_requests.
Jason Phang's avatar
checkin  
Jason Phang committed
601
        """
Leo Gao's avatar
Leo Gao committed
602
        pass
Jason Phang's avatar
gpt3  
Jason Phang committed
603

604
    @abstractmethod
605
606
    def aggregation(self):
        """
&'s avatar
&amp; committed
607
        :returns: {str: [metric_score] -> float}
Fabrizio Milo's avatar
Fabrizio Milo committed
608
            A dictionary where keys are the names of submetrics and values are
&'s avatar
&amp; committed
609
            functions that aggregate a list of metric scores
610
611
612
        """
        pass

613
    @abstractmethod
614
615
616
    def higher_is_better(self):
        """
        :returns: {str: bool}
Fabrizio Milo's avatar
Fabrizio Milo committed
617
            A dictionary where keys are the names of submetrics and values are
618
619
620
621
            whether a higher value of the submetric is better
        """
        pass

Jason Phang's avatar
Jason Phang committed
622
    def fewshot_description(self):
623
        import warnings
Fabrizio Milo's avatar
Fabrizio Milo committed
624

625
        warnings.warn(
Jonathan Tow's avatar
Jonathan Tow committed
626
            "`fewshot_description` will be removed in futures versions. Pass "
627
            "any custom descriptions to the `evaluate` function instead.",
Fabrizio Milo's avatar
Fabrizio Milo committed
628
629
            DeprecationWarning,
        )
Jason Phang's avatar
checkin  
Jason Phang committed
630
631
        return ""

632
    @utils.positional_deprecated
Fabrizio Milo's avatar
Fabrizio Milo committed
633
634
635
636
    def fewshot_context(
        self, doc, num_fewshot, provide_description=None, rnd=None, description=None
    ):
        """Returns a fewshot context string that is made up of a prepended description
637
638
639
640
641
642
643
644
645
646
        (if provided), the `num_fewshot` number of examples, and an appended prompt example.

        :param doc: str
            The document as returned from training_docs, validation_docs, or test_docs.
        :param num_fewshot: int
            The number of fewshot examples to provide in the returned context string.
        :param provide_description: bool
            Not implemented, and this option is deprecated and will be removed in a future version in favor of a different description providing method
        :param rnd: random.Random
            The pseudo-random number generator used to randomly sample examples.
647
            WARNING: This is currently a required arg although it's optionalized with a default `None`.
648
649
650
651
652
        :param description: str
            The task's description that will be prepended to the fewshot examples.
        :returns: str
            The fewshot context.
        """
Fabrizio Milo's avatar
Fabrizio Milo committed
653
654
655
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`"
Jonathan Tow's avatar
Merge  
Jonathan Tow committed
656
        assert not provide_description, (
Jonathan Tow's avatar
Jonathan Tow committed
657
            "The `provide_description` arg will be removed in future versions. To prepend "
658
            "a custom description to the context, supply the corresponding string via the "
Jonathan Tow's avatar
Jonathan Tow committed
659
            "`description` arg."
Jonathan Tow's avatar
Merge  
Jonathan Tow committed
660
        )
661
662
        if provide_description is not None:
            # nudge people to not specify it at all
Fabrizio Milo's avatar
Fabrizio Milo committed
663
664
665
            print(
                "WARNING: provide_description is deprecated and will be removed in a future version in favor of description_dict"
            )
666

667
        description = description + "\n\n" if description else ""
668

669
670
        if num_fewshot == 0:
            labeled_examples = ""
671
        else:
672
673
674
675
676
            # for sets with no training docs, draw from other set *but ensure no overlap with current doc*
            if self.has_training_docs():
                fewshotex = self.fewshot_examples(k=num_fewshot, rnd=rnd)
            else:
                if self._fewshot_docs is None:
677
                    self._fewshot_docs = list(
Fabrizio Milo's avatar
Fabrizio Milo committed
678
679
680
                        self.validation_docs()
                        if self.has_validation_docs()
                        else self.test_docs()
681
                    )
682

683
                fewshotex = rnd.sample(self._fewshot_docs, num_fewshot + 1)
684

685
686
                # get rid of the doc that's the one we're evaluating, if it's in the fewshot
                fewshotex = [x for x in fewshotex if x != doc][:num_fewshot]
687

Fabrizio Milo's avatar
Fabrizio Milo committed
688
689
690
691
692
693
694
695
696
            labeled_examples = (
                "\n\n".join(
                    [
                        self.doc_to_text(doc) + self.doc_to_target(doc)
                        for doc in fewshotex
                    ]
                )
                + "\n\n"
            )
Leo Gao's avatar
Update  
Leo Gao committed
697

698
        example = self.doc_to_text(doc)
Leo Gao's avatar
Leo Gao committed
699
700
701
        return description + labeled_examples + example


Jon Tow's avatar
Jon Tow committed
702
class MultipleChoiceTask(Task):
Leo Gao's avatar
Leo Gao committed
703
    def doc_to_target(self, doc):
Fabrizio Milo's avatar
Fabrizio Milo committed
704
        return " " + doc["choices"][doc["gold"]]
Leo Gao's avatar
Leo Gao committed
705

Leo Gao's avatar
Leo Gao committed
706
707
    def construct_requests(self, doc, ctx):
        lls = [
Fabrizio Milo's avatar
Fabrizio Milo committed
708
            rf.loglikelihood(ctx, " {}".format(choice))[0] for choice in doc["choices"]
Leo Gao's avatar
Leo Gao committed
709
710
711
712
713
714
715
        ]

        return lls

    def process_results(self, doc, results):
        gold = doc["gold"]

Fabrizio Milo's avatar
Fabrizio Milo committed
716
        acc = 1.0 if np.argmax(results) == gold else 0.0
717
        completion_len = np.array([float(len(i)) for i in doc["choices"]])
Fabrizio Milo's avatar
Fabrizio Milo committed
718
        acc_norm = 1.0 if np.argmax(results / completion_len) == gold else 0.0
Leo Gao's avatar
Leo Gao committed
719
720

        return {
Leo Gao's avatar
Leo Gao committed
721
722
            "acc": acc,
            "acc_norm": acc_norm,
Leo Gao's avatar
Leo Gao committed
723
        }
Fabrizio Milo's avatar
Fabrizio Milo committed
724

Leo Gao's avatar
Leo Gao committed
725
726
    def higher_is_better(self):
        return {
Leo Gao's avatar
Leo Gao committed
727
728
            "acc": True,
            "acc_norm": True,
Leo Gao's avatar
Leo Gao committed
729
        }
Fabrizio Milo's avatar
Fabrizio Milo committed
730

Leo Gao's avatar
Leo Gao committed
731
732
    def aggregation(self):
        return {
Leo Gao's avatar
Leo Gao committed
733
734
            "acc": mean,
            "acc_norm": mean,
Leo Gao's avatar
Leo Gao committed
735
736
737
        }


Jason Phang's avatar
Jason Phang committed
738
class PerplexityTask(Task, abc.ABC):
739
740
741
742
    def should_decontaminate(self):
        """Whether this task supports decontamination against model training set."""
        return True

Jason Phang's avatar
Jason Phang committed
743
744
745
746
747
748
749
    def has_training_docs(self):
        return False

    def fewshot_examples(self, k, rnd):
        assert k == 0
        return []

Fabrizio Milo's avatar
Fabrizio Milo committed
750
751
752
753
754
755
756
757
758
    def fewshot_context(
        self, doc, num_fewshot, provide_description=None, rnd=None, description=None
    ):
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`."
Jonathan Tow's avatar
Merge  
Jonathan Tow committed
759
        assert not provide_description, (
Jonathan Tow's avatar
Jonathan Tow committed
760
            "The `provide_description` arg will be removed in future versions. To prepend "
Jonathan Tow's avatar
Jonathan Tow committed
761
            "a custom description to the context, supply the corresponding string via the "
Jonathan Tow's avatar
Jonathan Tow committed
762
            "`description` arg."
Jonathan Tow's avatar
Merge  
Jonathan Tow committed
763
        )
764
765
        if provide_description is not None:
            # nudge people to not specify it at all
Fabrizio Milo's avatar
Fabrizio Milo committed
766
767
768
            print(
                "WARNING: provide_description is deprecated and will be removed in a future version in favor of description_dict"
            )
769

Jason Phang's avatar
Jason Phang committed
770
771
772
        return ""

    def higher_is_better(self):
Leo Gao's avatar
Leo Gao committed
773
774
775
776
777
        return {
            "word_perplexity": False,
            "byte_perplexity": False,
            "bits_per_byte": False,
        }
Jason Phang's avatar
Jason Phang committed
778

779
780
781
    def doc_to_decontamination_query(self, doc):
        return doc

Jason Phang's avatar
Jason Phang committed
782
    def doc_to_text(self, doc):
783
        return ""
Jason Phang's avatar
Jason Phang committed
784
785

    def doc_to_target(self, doc):
786
        return doc
Jason Phang's avatar
Jason Phang committed
787
788
789

    def construct_requests(self, doc, ctx):
        assert not ctx
Leo Gao's avatar
Leo Gao committed
790
        req = rf.loglikelihood_rolling(self.doc_to_target(doc))
Jason Phang's avatar
Jason Phang committed
791
792
793
        return req

    def process_results(self, doc, results):
Fabrizio Milo's avatar
Fabrizio Milo committed
794
        (loglikelihood,) = results
Leo Gao's avatar
Leo Gao committed
795
        words = self.count_words(doc)
796
        bytes_ = self.count_bytes(doc)
Jason Phang's avatar
Jason Phang committed
797
        return {
Leo Gao's avatar
Leo Gao committed
798
            "word_perplexity": (loglikelihood, words),
799
            "byte_perplexity": (loglikelihood, bytes_),
800
            "bits_per_byte": (loglikelihood, bytes_),
Jason Phang's avatar
Jason Phang committed
801
802
803
804
        }

    def aggregation(self):
        return {
Leo Gao's avatar
Leo Gao committed
805
806
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
807
            "bits_per_byte": bits_per_byte,
Jason Phang's avatar
Jason Phang committed
808
809
        }

810
811
    @classmethod
    def count_bytes(cls, doc):
Leo Gao's avatar
Leo Gao committed
812
        return len(doc.encode("utf-8"))
813
814
815

    @classmethod
    def count_words(cls, doc):
Fabrizio Milo's avatar
Fabrizio Milo committed
816
        """Downstream tasks with custom word boundaries should override this!"""
Leo Gao's avatar
Leo Gao committed
817
        return len(re.split(r"\s+", doc))
Leo Gao's avatar
Leo Gao committed
818

Jason Phang's avatar
Jason Phang committed
819

Leo Gao's avatar
Leo Gao committed
820
821
def hash_args(attr, args):
    dat = json.dumps([attr] + list(args))
Fabrizio Milo's avatar
Fabrizio Milo committed
822
    return hashlib.sha256(dat.encode("utf-8")).hexdigest()
Leo Gao's avatar
Leo Gao committed
823
824


Leo Gao's avatar
Leo Gao committed
825
826
class CacheHook:
    def __init__(self, cachinglm):
Fabrizio Milo's avatar
Fabrizio Milo committed
827
        if cachinglm is None:
Leo Gao's avatar
Leo Gao committed
828
829
830
831
            self.dbdict = None
            return

        self.dbdict = cachinglm.dbdict
Fabrizio Milo's avatar
Fabrizio Milo committed
832

Leo Gao's avatar
Leo Gao committed
833
834
835
836
837
838
839
    def add_partial(self, attr, req, res):
        if self.dbdict is None:
            return
        hsh = hash_args(attr, req)
        self.dbdict[hsh] = res


Leo Gao's avatar
Leo Gao committed
840
841
class CachingLM:
    def __init__(self, lm, cache_db):
842
843
844
845
846
847
848
        """LM wrapper that returns cached results if they exist, and uses the underlying LM if not.

        :param lm: LM
            Underlying LM
        :param cache_db: str
            Path to cache db
        """
Leo Gao's avatar
Leo Gao committed
849
850
        self.lm = lm
        self.cache_db = cache_db
851
852
        if os.path.dirname(cache_db):
            os.makedirs(os.path.dirname(cache_db), exist_ok=True)
Leo Gao's avatar
Leo Gao committed
853
854
        self.dbdict = SqliteDict(cache_db, autocommit=True)

Leo Gao's avatar
Leo Gao committed
855
856
857
        # add hook to lm
        lm.set_cache_hook(self.get_cache_hook())

Leo Gao's avatar
Leo Gao committed
858
859
860
861
    def __getattr__(self, attr):
        def fn(requests):
            res = []
            remaining_reqs = []
Fabrizio Milo's avatar
Fabrizio Milo committed
862

Leo Gao's avatar
Leo Gao committed
863
864
            # figure out which ones are cached and which ones are new
            for req in requests:
Leo Gao's avatar
Leo Gao committed
865
                hsh = hash_args(attr, req)
Leo Gao's avatar
Leo Gao committed
866
867
868
869
870
871
872
873
874
                if hsh in self.dbdict:
                    ob = self.dbdict[hsh]

                    assert ob is not None

                    res.append(ob)
                else:
                    res.append(None)
                    remaining_reqs.append(req)
Fabrizio Milo's avatar
Fabrizio Milo committed
875

876
            # actually run the LM on the requests that do not have cached results
Leo Gao's avatar
Leo Gao committed
877
878
879
880
881
            rem_res = getattr(self.lm, attr)(remaining_reqs)

            # stick the new ones back into the list and also cache any of the new ones
            resptr = 0
            for req, r in zip(remaining_reqs, rem_res):
882
883
                while res[resptr] is not None:
                    resptr += 1
Leo Gao's avatar
Leo Gao committed
884
885
886
887

                res[resptr] = r

                # caching
Leo Gao's avatar
Leo Gao committed
888
                hsh = hash_args(attr, req)
Leo Gao's avatar
Leo Gao committed
889
                self.dbdict[hsh] = r
Leo Gao's avatar
Leo Gao committed
890
            self.dbdict.commit()
Leo Gao's avatar
Leo Gao committed
891
892

            return res
Fabrizio Milo's avatar
Fabrizio Milo committed
893

Leo Gao's avatar
Leo Gao committed
894
        return fn
Fabrizio Milo's avatar
Fabrizio Milo committed
895

Leo Gao's avatar
Leo Gao committed
896
897
    def get_cache_hook(self):
        return CacheHook(self)
Leo Gao's avatar
Leo Gao committed
898

Jason Phang's avatar
Jason Phang committed
899

900
REQUEST_RETURN_LENGTHS = {
Fabrizio Milo's avatar
Fabrizio Milo committed
901
902
903
    "loglikelihood": 2,
    "greedy_until": None,
    "loglikelihood_rolling": None,
904
905
906
}


907
class Request:
Leo Gao's avatar
Leo Gao committed
908
909
    def __init__(self, request_type, args, index=None):
        if request_type not in REQUEST_RETURN_LENGTHS.keys():
Fabrizio Milo's avatar
Fabrizio Milo committed
910
911
912
            raise NotImplementedError(
                "The request type {} is not implemented!".format(request_type)
            )
Leo Gao's avatar
Leo Gao committed
913

Leo Gao's avatar
Leo Gao committed
914
        self.request_type = request_type
915
916
        self.args = args
        self.index = index
Fabrizio Milo's avatar
Fabrizio Milo committed
917

918
    def __iter__(self):
Leo Gao's avatar
Leo Gao committed
919
        if REQUEST_RETURN_LENGTHS[self.request_type] is None:
Fabrizio Milo's avatar
Fabrizio Milo committed
920
            raise IndexError("This request type does not return multiple arguments!")
Leo Gao's avatar
Leo Gao committed
921
922
        for i in range(REQUEST_RETURN_LENGTHS[self.request_type]):
            yield Request(self.request_type, self.args, i)
Fabrizio Milo's avatar
Fabrizio Milo committed
923

924
    def __getitem__(self, i):
Leo Gao's avatar
Leo Gao committed
925
        if REQUEST_RETURN_LENGTHS[self.request_type] is None:
Fabrizio Milo's avatar
Fabrizio Milo committed
926
            raise IndexError("This request type does not return multiple arguments!")
Leo Gao's avatar
Leo Gao committed
927
        return Request(self.request_type, self.args, i)
Fabrizio Milo's avatar
Fabrizio Milo committed
928

Leo Gao's avatar
Leo Gao committed
929
    def __eq__(self, other):
Fabrizio Milo's avatar
Fabrizio Milo committed
930
931
932
933
934
        return (
            self.request_type == other.request_type
            and self.args == other.args
            and self.index == other.index
        )
Leo Gao's avatar
Leo Gao committed
935

Leo Gao's avatar
Leo Gao committed
936
    def __repr__(self):
Leo Gao's avatar
Leo Gao committed
937
        return f"Req_{self.request_type}{self.args}[{self.index}]\n"
938

Jason Phang's avatar
Jason Phang committed
939

Leo Gao's avatar
Leo Gao committed
940
941
class RequestFactory:
    def __getattr__(self, attr):
Leo Gao's avatar
Update  
Leo Gao committed
942
943
        def fn(*args):
            return Request(attr, args)
Fabrizio Milo's avatar
Fabrizio Milo committed
944

Leo Gao's avatar
Leo Gao committed
945
946
947
948
        return fn


rf = RequestFactory()